Deep Learning Enables Instant and Versatile Estimation of Rice Yield Using Ground-Based RGB Images.

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Rice (Oryza sativa L.) is a vital global food source, but accurate yield assessment, especially in the Global South, remains challenging.
• High-throughput phenotyping is crucial for improving crop productivity and food security.

Purpose of the Study:

• To develop and validate a deep-learning-based approach for instantaneous rice yield estimation using red-green-blue (RGB) images.
• To assess the model's robustness and scalability for practical agricultural applications.

Main Methods:

• A convolutional neural network (CNN) was trained on over 22,000 RGB images captured from 4,820 rice plots across Africa and Japan.
• The model was evaluated for its accuracy in predicting yield variation, genotypic differences, and the impact of agronomic interventions.

Main Results:

• The CNN model explained 68% of the yield variation with a relative root mean square error of 0.22.
• The model demonstrated robustness to variations in shooting angles, lighting, and image resolution, predicting 57% of yield variation even with reduced resolution.
• The approach successfully identified genotypic differences and the effects of agronomic practices on rice yield.

Conclusions:

• Deep learning applied to RGB imagery offers a low-cost, rapid, and scalable method for high-throughput rice phenotyping and yield estimation.
• This technology can aid in assessing productivity-enhancing interventions, identifying areas needing intervention, and forecasting yield weeks before harvest.